Vehicle Underside Inspection Using GAN-Generated Defect Images
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing vehicle inspection methods for the lower side, which rely on manual visual inspection, lead to operator fatigue and inefficiencies due to low defect rates, making it difficult to train AI models effectively for defect detection.
Innovation Solution
A method and apparatus using a generative adversarial network (GAN) to generate defect images from virtual images, combined with a second learning model for accurate defect detection, enabling efficient AI-based inspection of vehicle lower sides.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If manual visual inspection is used to inspect the lower side of the vehicle, then the inspection can be performed without AI model training, but operator fatigue increases and inspection efficiency decreases
Solution Approach 1:
The patent replaces manual visual inspection with an automated AI-based image recognition system. A camera captures images of the vehicle's lower side, and a pre-trained AI model automatically detects defects, eliminating the need for operators to manually inspect and thereby resolving the contradiction between ease of implementation and inspection efficiency.
2Measurement precision
If AI model training is performed using actual defect data from vehicle lower sides, then defect detection accuracy can be improved, but the low defect rate makes it difficult to obtain sufficient training data
Solution Approach 1:
The patent uses a data augmentation technique where a small set of actual defect images are copied and transformed through various operations (rotation, flipping, cropping, color adjustments) to generate a large volume of diverse training data. This resolves the contradiction by enabling sufficient training data quantity while maintaining defect detection accuracy through the use of real defect patterns.
Solution Approach 2:
The patent applies parameter transformations to defect images including rotation angles, flipping directions, cropping regions, and color space adjustments. These parameter changes create varied training samples from limited original defect data, solving the data scarcity problem while preserving the essential defect characteristics for accurate detection.
3Quantity of substance
If more defect images are generated through data augmentation, then training data quantity increases, but the complexity of the data processing pipeline increases
Solution Approach 1:
The data processing pipeline is segmented into distinct, modular operations: image loading, rotation, flipping, cropping, color adjustment, and saving. Each operation is implemented as a separate function that can be independently configured and executed. This segmentation reduces complexity by making the processing pipeline manageable, testable, and easily adjustable while still generating abundant training data.
Data Source
AI summary
An apparatus and method for inspecting a vehicle include photographing a lower side of a vehicle to obtain a lower side image, transmitting the lower image to a server, receiving inspection result data of the lower side of the vehicle from the server, and displaying the inspection result data on a display, wherein the received inspection result data includes a result obtained by generating a defect image of the lower side of the vehicle from a virtual image, in which a defect is indicated, of a reference lower image of the vehicle, and determine whether the lower side of the vehicle is defective using the generated defect image.